impl thread compaction

1 files changed, 41 insertions, 0 deletions

diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index ed91a44..deebfd6 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -88,6 +88,47 @@ void warp_inst_t::compute_simd_sets(unsigned num_sets, unsigned set_width) {
   }
 }
 
+void warp_inst_t::compute_simd_sets_compacted(unsigned num_sets,
+                                               unsigned set_width,
+                                               unsigned start_set) {
+  m_simd_sets.resize(num_sets);
+  // Initialize all sets as invalid
+  for (unsigned s = 0; s < num_sets; s++) {
+    simd_set_info &info = m_simd_sets[s];
+    info.set_id = s;
+    info.warp_id = m_warp_id;
+    info.set_active_mask.reset();
+    info.active_mask_in_warp.reset();
+    info.num_active_threads = 0;
+    info.valid = false;
+    info.source_inst = NULL;
+    for (unsigned i = 0; i < MAX_WARP_SIZE; i++) info.thread_ids[i] = 0;
+  }
+
+  // Collect active thread IDs
+  std::vector<unsigned> active_tids;
+  for (unsigned t = 0; t < m_config->warp_size; t++) {
+    if (m_warp_active_mask.test(t)) active_tids.push_back(t);
+  }
+
+  // Pack active threads contiguously into sets starting at start_set
+  unsigned tid_idx = 0;
+  for (unsigned s = start_set; s < num_sets && tid_idx < active_tids.size();
+       s++) {
+    simd_set_info &info = m_simd_sets[s];
+    for (unsigned lane = 0;
+         lane < set_width && tid_idx < active_tids.size(); lane++) {
+      unsigned tid = active_tids[tid_idx];
+      info.thread_ids[lane] = tid;
+      info.set_active_mask.set(lane);
+      info.active_mask_in_warp.set(tid);
+      info.num_active_threads++;
+      tid_idx++;
+    }
+    info.valid = true;
+  }
+}
+
 unsigned warp_inst_t::num_active_simd_sets() const {
   unsigned count = 0;
   for (unsigned i = 0; i < m_simd_sets.size(); i++) {